Multi-Ship Future Interaction Trajectory Prediction via Pre-Initializer Diffusion Model

Real-time stochastic multi-ship trajectory modeling is crucial for maritime safety. However, it remains challenging due to the uncertainty of dynamic vessel intentions and their complex interactions. Most existing studies rely on deterministic social data from historical time steps for modeling, which often fail to capture the future states of interacting ships, leading to unrealistic trajectory overlaps. Recent research has demonstrated that diffusion models excel in trajectory prediction due to their high generation quality, training stability, and diversity. However, their slow sampling speed limits real-time perception in maritime environments, as generating high-quality trajectories typically requires hundreds of denoising steps. To address these challenges, we propose a Multi-Ship Future interaction trajectory prediction approach based on a Pre-initializer Diffusion model (MFPD). By training a parameterized pre-initializer to directly learn the joint distribution of multiple denoising steps in the reverse diffusion process, our method significantly reduces the time cost of denoising while retaining only a few steps for fine-tuning the distribution. Specifically, in addition to encoding historical trajectory information and social interactions as state embeddings, we also incorporate future trajectory and multimodal maritime environmental information as input condition embeddings to fully capture potential future interactions and environmental features. Experimental results demonstrate that the proposed model significantly improves performance on two real-world datasets while greatly accelerating the sampling speed, demonstrating the superiority in real-world maritime environments.

Authors

• Kun Ma
• Qilong Han Harbin Engineering University
• Jingzheng Yao Harbin Engineering University

Keywords

• Ship Automatic Identification System
• Multi-ship
• Trajectory Prediction
• Diffusion Models
• Parameterized Pre-initializer